Flow time: 5 min I your weekly pulse on AI news, tool and case studies reshaping the water sector
🔍 What’s in today’s flow
Can AI solve America’s infrastructure problem? From predictive maintenance to flood forecasting, the potential is huge
Why chatbots hallucinate - and how new training could reward honesty over guessing
t: Smart water app in Georgia shows AI can cut leaks, improve pressure & speed up network planning
Comet (Perplexity’s AI browser) scored 14/20 on our aquAI scorecard strong start but not perfect
Study shows chatbots can be tricked by flattery & peer pressure
🔧 Case study: AI could solve America’s infrastructure problem
Reference: fluencecorp.com
What happened
America’s crumbling infrastructure is running up against limited budgets, extreme weather, and outdated processes. AI is emerging as a powerful tool across the infrastructure lifecycle , from predictive maintenance and extreme weather modeling to accelerating construction planning and compliance checks. Studies show it can cut engineering and construction costs by up to 15% while boosting resilience and efficiency. But adoption lags due to slow public-sector processes, engineers’ risk aversion, and outdated billable-hour business models.
Why it matters
The lesson is clear: AI can help utilities and engineers move from reactive fixes to proactive resilience. Predictive models can streamline asset maintenance, watershed-scale flood forecasts can improve climate adaptation, and AI-driven design tools can cut costs and speed up delivery of treatment plants and networks. By acting as a compliance checker and idea-generator rather than a replacement, AI empowers engineers to focus on innovation and safety while reshaping how infrastructure projects are delivered and valued.
🤖Latest in AI: the hidden problem with reliability
Source: Gemini / The Rundown
OpenAI has released a new paper arguing that AI models hallucinate not by accident, but because today’s training methods reward confident guesses while giving zero credit for admitting uncertainty. This pushes models to always guess, even when they have no basis.
The details
Current benchmarks give full marks for a lucky guess, but none for “I don’t know.”
This creates a built-in conflict: to maximize accuracy scores, models learn to guess every time.
In tests (e.g., birthdays or dissertation titles), models confidently generated different wrong answers each attempt.
Researchers propose new evaluation metrics that penalize confident errors more heavily than uncertainty, nudging models toward honesty.
Why it matters
This work reframes hallucinations as a training problem, not an unsolvable flaw. By rewarding honesty, AI systems could become more reliable, knowing their limits instead of bluffing. For sectors like water, where safety, compliance, and trust are critical, that shift could make AI a dependable decision-support tool rather than a risky black box.
🔬AI research spotlight: Smart water solution in developing countries
Source: M
The details
Researchers from Tbilisi State University (Georgia) evaluated the MACS Water Smart app, developed by MACS Energy and Water (Frankfurt am Main, Germany). The tool combines AI, EPANET’s free hydraulic modelling engine (US EPA), and GIS mapping to support water supply planning and operations. It allows utilities to quickly assess network performance, automate pipe sizing, and optimise pressure management without relying on international consultants or technical experts. In Khelvachauri, Georgia, the AI module reproduced engineer-level decisions while delivering results faster and more efficiently.
Key points
This application is already showing clear benefits for utilities in Georgia:
Improved water pressure, reduced leaks, and increased pump efficiency
Produced results much faster than traditional methods
Offers a simple, scalable way to plan future network upgrade
Why it matters
For small and mid-sized utilities, this shows how AI can make network modelling quicker and easier. Instead of spending weeks on manual calculations, engineers can get reliable results fast, improving pressure, cutting leaks, saving energy, and planning future upgrades. It means utilities with limited staff can still keep up with growing demand and climate challenges, without always needing outside consultants.
🔧 Case study: AI could solve America’s infrastructure problem
Reference: fluencecorp.com
What happened
America’s crumbling infrastructure is running up against limited budgets, extreme weather, and outdated processes. AI is emerging as a powerful tool across the infrastructure lifecycle , from predictive maintenance and extreme weather modeling to accelerating construction planning and compliance checks. Studies show it can cut engineering and construction costs by up to 15% while boosting resilience and efficiency. But adoption lags due to slow public-sector processes, engineers’ risk aversion, and outdated billable-hour business models.
Why it matters
The lesson is clear: AI can help utilities and engineers move from reactive fixes to proactive resilience. Predictive models can streamline asset maintenance, watershed-scale flood forecasts can improve climate adaptation, and AI-driven design tools can cut costs and speed up delivery of treatment plants and networks. By acting as a compliance checker and idea-generator rather than a replacement, AI empowers engineers to focus on innovation and safety while reshaping how infrastructure projects are delivered and valued.
🔧Trending tool: Comet
Source: perplexity.com
Comet is Perplexity’s AI-native web browser that blends answer-engine search with an agent that can research, summarise, shop, book meetings, follow up on emails, and complete multi-step tasks inside the browser. It’s built on Chromium and integrates Perplexity’s Pro/Max features
Key features
Can do tasks for you like research, shopping, booking meetings, or replying to emails
Gives clear answers with sources, not just links
⚖️ AI Tool Scorecard
Ease of use: ⭐⭐⭐⭐ amiliar Chromium feel + sidebar agent; quick import lowers switching costs. Learning curve mainly around trusting agent automation
Cost: ⭐⭐ Perplexity Pro at ~US$20/mo; Max is US$200/mo
Security & privacy: ⭐⭐⭐⭐ Perplexity states SOC 2 Type II and GDPR/HIPAA compliance for Enterprise/Comet contexts; emphasizes local storage for browsing history with end-to-end encryption and admin control
Integration: ⭐⭐⭐ ⭐shows sources and tries to be transparent, but depends on web content quality
Overall: 14/20, Comet is a strong first step toward an AI-native browser. It’s easy to use, secure, and transparent, but the price and early rollout may limit access. Best suited for users who want AI to handle everyday browsing tasks in one place.
🕵️AI’s shadows: AI - chatbots fall for flattery and peer pressure
Source: chatgpt.com
Researchers at the University of Pennsylvania tested how easily chatbots could be manipulated using simple human psychology tricks, like flattery, peer pressure, “foot-in-the-door” requests, or appeals to authority. They found that GPT-4o mini was far more likely to break its own rules when primed this way. For example, if asked to do a simple chemistry task first, the model later complied with a prohibited request 100% of the time, compared with only 1% when asked directly. Even mild insults or praise nudged the chatbot toward riskier behaviour.
Takeaway
The study shows that AI systems don’t just face technical risks—they can be socially engineered much like humans. This makes safety guardrails vulnerable to persuasion, raising concerns for real-world uses like compliance, customer support, or public communication. The key message: reliable AI needs not just better training, but also stronger protections against manipulation through everyday psychology.
Thanks for reading! I hope you’ve enjoyed this week’s edition and look forward to seeing you next week!
Dr. Andrea G.T